Lightweight grape leaf disease recognition method based on transformer framework.

Area of Science:

• Agricultural Science
• Computer Vision
• Machine Learning

Background:

• Accurate grape disease recognition is vital for preventing yield loss.
• Small-sample conditions pose challenges in feature extraction and data augmentation for grape leaf disease detection.

Purpose of the Study:

• To address limitations in small-sample grape leaf disease recognition.
• To propose a novel deep learning approach combining multi-scale feature fusion and generative data augmentation.

Main Methods:

• Developed a multi-scale feature hybrid fusion architecture with enhanced data augmentation.
• Introduced the LVT Block (Ghost and Transformer) for multi-scale and global information perception.
• Proposed the DLVT Block by combining LVT and MARI blocks for richer feature representation and disease area perception, forming the DLVTNet model.

Main Results:

• Achieved an average recognition rate of 98.48% on the New Plant Diseases Dataset.
• Reduced model parameters to 42.7% of MobileNetV4 while maintaining high accuracy.
• Demonstrated 96.12% accuracy in tomato leaf disease testing, indicating strong generalization.

Conclusions:

• The proposed DLVTNet effectively alleviates sample insufficiency in intelligent agricultural detection.
• The method offers a new system for disease detection with strong interpretability and excellent generalization performance.